Yesterday, Meituan released its first open-source large model: LongCat-Flash, a 560-billion-parameter mixture-of-experts model whose signature trait is speed. Meituan says its shortcut-connected architecture extends the compute-communication overlap window to deliver more than 100 tokens per second, with theoretical per-token output time nearly 50 percent lower than DeepSeek-V3. Zhidx’s hands-on test confirmed it: requesting roughly 100 tokens of output, the measured generation speed indeed ran around 100 tokens per second after first-token latency.
Caveats first: LongCat-Flash does not yet support file or image uploads; the web version offers web search, while deep thinking mode shows “coming soon.” And on a May earnings call, founder and CEO Wang Xing had already previewed the direction — the enhanced LongCat “can now switch seamlessly between reasoning and non-reasoning modes,” with performance approaching GPT-4o — while setting the strategic tone: Meituan plays offense in AI, spanning AI at work, AI in products, and self-developed LLMs. Per 01 Founder’s reporting, cross-confirmed with multiple sources, LongCat-Flash’s training was not completed on NVIDIA GPUs — the technical report says only “tens of thousands of accelerator cards,” with the hardware vendor’s name deliberately undisclosed.
Hands-On: Sharp Commentary, Failed Elementary Math
Zhidx tested math, coding, search, and general knowledge. Knowledge cutoff is June 2024, and it shows offline: asked who develops Tencent Yuanbao, the model answered Meituan — Yuanbao launched May 30, 2024, right at the cutoff boundary. Given the common agent workflow of searching, synthesizing, and producing a research report on the latest “food delivery war” among Meituan, Alibaba, and JD, LongCat-Flash searched only 10 pages — fewer than peers — and its sources skewed heavily toward Baijiahao, missing authoritative media and company official sites even when explicitly requested. The cited financial figures were accurate and the analysis matched mainstream industry reads, favoring Alibaba — but offered little new insight. On the classic 9.8 versus 9.11 comparison, it stumbled: a fast, confidently wrong answer with no self-verification. On more contextualized math (interest calculations) it performed fully, and on general knowledge — why the sky is blue at noon and red at sunset — it delivered a three-part explanation with correct theory, a clean table, even suggested popular-science references. In coding, a LongCat-themed HTML mini-game worked but played poorly. Against DeepSeek-V3.1, Kimi-K2, and Qwen3 MoE-2507, general and coding ability still trails — but in agentic tool calling and instruction following, LongCat-Flash reaches open-source SOTA.
The Novel MoE: A Shortcut for Expert Cooperation
The technical report details the architecture behind the speed. Zero-computation experts: next-token prediction carries wildly varying difficulty — hard tokens (complex reasoning, ambiguous context) need full compute; easy tokens (common words, formulaic structures) barely need any. LongCat-Flash’s dynamic compute-allocation mechanism routes some simple tokens to zero-computation experts that pass input through unchanged, costing nothing. Per-token expert activation varies with context, so the model spends more compute on important tokens and less on trivial ones — better performance within the same budget. Per-expert bias terms adjust routing scores based on recent usage, keeping the zero-computation share balanced without touching the language-model objective. Shortcut-Connected MoE (ScMoE): at scale, communication is the bottleneck — traditional expert parallelism must complete global routing before computing begins. Shared-expert architectures overlap communication with only a single expert’s compute, which is too narrow to help much. ScMoE introduces a cross-layer shortcut that reorders execution: the previous layer’s dense computation runs in parallel with the current MoE layer’s dispatch/aggregation communication — a much larger overlap window. Training-loss curves match non-ScMoE baselines, so performance is unharmed, while inference enables a single-batch overlapped pipeline driving the ~50 percent per-token output time reduction. It even parallelizes different communication modes: intra-node tensor parallelism over NVLink fully overlaps inter-node expert parallelism over RDMA. Scalable variance alignment: architectures that behave at small scale can destabilize at large scale, and the team identified mismatched variance inside modules as the culprit — fixing it with two scale-correction factors plus variance-compensated expert initialization for MLA and MoE blocks.
Scaling From a Half-Size Model, With a Stability Framework
The training strategy is a complete stability-and-scalability framework. Rather than trial-and-error at hundreds of billions of parameters, the team ran a “small-to-large” strategy: systematically search learning rates, initialization variances, and other hyperparameters on a small model, then migrate the optimal configurations through theoretically derived rules — saving enormous compute while keeping hyperparameters valid at scale. Initialization used model growth: train a half-scale model to convergence, then expand it to full size and continue — brief early wobble, then stable convergence, clearly outperforming random initialization. Stability stack: routing-level load balancing across experts; an activation “fuse” that gently constrains anomalous spikes to prevent training collapse; and carefully tuned Adam parameters for numerical stability at scale — together eliminating unrecoverable loss spikes across long runs. Reliability work emphasized reproducibility and error detection, with strict precision control catching “silent data corruption” before it hides inside distributed systems. The training pipeline itself: two-stage pretraining data fusion that progressively raises STEM-and-code share; mid-training extending context to 128K tokens with synthetic data strengthening reasoning and coding; and post-training with separate tracks for reasoning, coding, and agent tasks — including a multi-agent synthetic framework generating hard tasks across information processing, tool complexity, and user interaction to train the agent capability. Infrastructure work covered kernel optimization, distributed parallelism, and monitoring/fault tolerance for tens of thousands of accelerator cards. Serving combines compute-communication orchestration, speculative decoding, KV-cache optimization, quantization, and scheduling: 100+ tokens per second at $0.70 per million tokens (about 5 RMB).
The Takeaway: Meituan Plays Offense
The team evaluated the model on 30 benchmarks: agent tasks are the long suit — a clear edge over DeepSeek-V3.1, Qwen3 MoE-2507, and Kimi-K2 on domestic open source — with coding, math, and reasoning still holding room to grow. The open-source goals, per the team: advancing efficient MoE architecture, high-quality data strategies, and autonomous agent model research. And per the earnings call, Meituan is already using AI to optimize existing products and build new AI-native ones — AI operations assistants, consumer AI assistants. The strategic reading: LongCat-Flash is not a me-too release. Its distinctive choices — agent-first post-training, non-NVIDIA training at scale, speed-over-marginal-benchmarks engineering, commodity pricing — sketch a lab that knows exactly which market it is entering (agents serving local-services scale) and is building for that market specifically. The LongCat-2.0 trillion-parameter follow-up covered earlier on this site now reads as the roadmap’s next station, and the open-source ecosystem gains a serious, unusually focused competitor. Weights on Hugging Face and GitHub with the technical report; web demo live.
